Asynchronous optimization for sequence training of neural networks
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for obtaining, by a first sequence-training speech model, a first batch of training frames that represent speech features of first training utterances; obtaining, by the first sequence-training speech model, one or more first neural network parameters; determining, by the first sequence-training speech model, one or more optimized first neural network parameters based on (i) the first batch of training frames and (ii) the one or more first neural network parameters; obtaining, by a second sequence-training speech model, a second batch of training frames that represent speech features of second training utterances; obtaining one or more second neural network parameters; and determining, by the second sequence-training speech model, one or more optimized second neural network parameters based on (i) the second batch of training frames and (ii) the one or more second neural network parameters.
1. A computer-implemented method executed on data processing hardware that causes the data processing hardware to perform operations comprising:
receiving a replica of a neural network model;
obtaining a batch of training utterances each comprising one or more predetermined words spoken by a speaker;
training, by performing stochastic gradient descent optimization on the batch of training utterances, the replica of the neural network model to generate corresponding model parameter gradients for the neural network model; and
sending the corresponding model parameter gradients for the neural network model to a centralized server.
2. The computer-implemented method of claim 1 , wherein sending the corresponding model parameter gradients for the neural network model comprises sending the corresponding model parameter gradients for the neural network model to the centralized network without sending the batch of training utterances to the centralized server.
3. The computer-implemented method of claim 1 , wherein:
the received replica of the neural network model comprises a current set of parameter values; and
training the replica of the neural network model comprises:
processing the batch of training utterances using the current set of parameter values for the neural network;
determining updated parameter values for the neural network model based on the processing of the batch of training utterances; and
performing stochastic gradient descent optimization to determine the corresponding model parameter gradients for the neural network model based on differences between the current set of parameter values and the updated parameter values for the neural network model.
4. The computer-implemented method of claim 3 , wherein the current set of parameter values for the neural network model comprise weights and biases of hidden layers of the neural network model.
5. The computer-implemented method of claim 4 , wherein the updated parameter values for the neural network model comprise updated weights for the neural network model.
6. The computer-implemented method of claim 1 , wherein the operations further comprise, after sending the corresponding model parameter gradients for the neural network to the centralized server, receiving an updated set of model parameters for the neural network model based on the corresponding model parameter gradients.
7. The computer-implemented method of claim 1 , wherein the neural network model is trained to indicate likelihoods that acoustic feature vectors represent different phonetic units.
8. The computer-implemented method of claim 1 , wherein the data processing hardware resides on a respective computing device associated with the speaker that spoke each training utterance in the batch of training utterances.
9. The computer-implemented method of claim 8 , wherein each obtained training utterances in the batch of training utterances is recorded by the respective computing device.
10. The computer-implemented method of claim 1 , wherein training the replica of the neural network model to generate corresponding model parameter gradients for the neural network model occurs in parallel with one or more data processing apparatuses training corresponding ones of other replicas of the neural network model.
11. A system comprising:
data processing hardware; and
memory hardware in communication with the data processing hardware and storing instructions that when executed on the data processing hardware cause the data processing hardware to perform operations comprising:
receiving a replica of a neural network model;
obtaining a batch of training utterances each comprising one or more predetermined words spoken by a speaker;
training, by performing stochastic gradient descent optimization on the batch of training utterances, the replica of the neural network model to generate corresponding model parameter gradients for the neural network model; and
sending the corresponding model parameter gradients for the neural network model to a centralized server.
12. The system of claim 11 , wherein sending the corresponding model parameter gradients for the neural network model comprises sending the corresponding model parameter gradients for the neural network model to the centralized network without sending the batch of training utterances to the centralized server.
13. The system of claim 11 , wherein:
the received replica of the neural network model comprises a current set of parameter values; and
training the replica of the neural network model comprises:
processing the batch of training utterances using the current set of parameter values for the neural network;
determining updated parameter values for the neural network model based on the processing of the batch of training utterances; and
performing stochastic gradient descent optimization to determine the corresponding model parameter gradients for the neural network model based on differences between the current set of parameter values and the updated parameter values for the neural network model.
14. The system of claim 13 , wherein the current set of parameter values for the neural network model comprise weights and biases of hidden layers of the neural network model.
15. The system of claim 14 , wherein the updated parameter values for the neural network model comprise updated weights for the neural network model.
16. The system of claim 11 , wherein the operations further comprise, after sending the corresponding model parameter gradients for the neural network to the centralized server, receiving an updated set of model parameters for the neural network model based on the corresponding model parameter gradients.
17. The system of claim 11 , wherein the neural network model is trained to indicate likelihoods that acoustic feature vectors represent different phonetic units.
18. The system of claim 11 , wherein the data processing hardware resides on a respective computing device associated with the speaker that spoke each training utterance in the batch of training utterances.
19. The system of claim 18 , wherein each obtained training utterances in the batch of training utterances is recorded by the respective computing device.
20. The system of claim 11 , wherein training the replica of the neural network model to generate corresponding model parameter gradients for the neural network model occurs in parallel with one or more data processing apparatuses training corresponding ones of other replicas of the neural network model.